Abstract Current approaches to AI alignment predominantly evaluate system behaviour at the level of individual outputs: correctness, normative compliance, preference satisfaction, or reward maximisation. This paper argues that a structurally distinct class of alignment failure exists that is invisible to output-level evaluation. We introduce the Parasitic Lock: a pathological interaction dynamic in which a human–AI system locally optimises task performance while inducing a net degradation of human cognitive agency over time. The defining characteristic is a directional transfer of interpretive burden and uncertainty resolution from the artificial system to the human participant, occurring within interaction sequences that remain instrumentally effective by all conventional measures. We situate the Parasitic Lock within a broader framework of ecological homeostasis, treating alignment not as a property of outputs but as a stability condition on joint human–AI system trajectories. The construct is introduced as a diagnostic criterion—not a prescriptive norm—allowing identification of failure modes without recourse to moral psychology, preference elicitation, or behavioural imitation. We argue that preventing such failures requires constraints that operate on reachable system states rather than on surface-level performance metrics. This paper is hypothetical and conceptual in nature. It proposes no implementation, makes no empirical claims, and introduces no hardware or energy-based models. Its contribution is a formal vocabulary and a structural argument for why the alignment problem cannot be solved at the interface layer alone. Keywords: AI alignment, human–AI interaction, cognitive agency, ecological homeostasis, system stability, parasitic dynamics, interaction failure modes